Latest / The Overlap / Silicon & Sparks: The Energy Cost of Crypto & AI (Part 1 of 2)

Transcript

Silicon & Sparks: The Energy Cost of Crypto & AI (Part 1 of 2) · The Overlap

• 0:00
  Welcome back. This week, we are plunging head first into the digital abyss. Abliss? No, abyss. We're plunging head first into the digital abyss. We're talking about the two things that are either going to save us all or just make everything a flaming dumpster fire. I'm talking, of course, about the seemingly unstoppable rise of cryptocurrency and the only technology that seems to be both our inevitable, unemployable demise and the truth.
• 0:29
  Savior of all mankind, artificial intelligence. We're going to yank back the curtain on what these things actually are, get into the gory details of their insane energy consumption, and then, of course, because we're gluttons for punishment, we're going to figure out if this whole mess could actually spark the next big energy revolution or, you know, just lead to a lot of really confident sounding computer generated technocracy. Let's find out.
• 0:58
  All righty. Welcome back, folks. I am your co-host Joshua. With me is your other favorite co-host, Will. Will, say hello.
• 1:23
  Hello. Good to be here with you, Joshua.
• 1:27
  Of course, as always, it's good to be here with you, too, and happy to have you back here for a second week, a second time this week.
• 1:35
  We have had quite a quite a busy week, both of productivity as well as of General malaise at what's going on in the world. So this is going to be a little bit of a I don't know, a bifurcation of that that seeming hellscape of constant news. So do you have an opinion? Have you done the research that I've done? Of course you have. I mean, but what do you got going on in this space?
• 2:04
  Do you have AI in the in the crypto space? Do you have crypto?
• 2:08
  I do have a bit of crypto. I've dipped my toe in the crypto. As they say, they might say. Yeah, I've got a little bit of that and I've heavily dipped into the AI. I'm very involved with AI. So there's a lot of investment, a lot of investment of time and focus in that area.
• 2:28
  So I mean, I think what what I've concluded is what you've also concluded that these are the the probably the defining technologies are at least a couple of the defining technologies of the 21st century.
• 2:41
  In some interesting ways, but specifically for our purposes this week, we're talking about the fact that they are both major, major energy sink like they are concerned consumptive, hungry, hungry technologies, energy, hungry technologies.
• 2:58
  And that's because they're both computationally intensive and computationally intensive is one of those words I don't want to say too many times, but that signals when I think computational intensity.
• 3:12
  I I think of Joshua. So I'm going to have you enlighten our audience on the technical side of that.
• 3:20
  But because both these technologies involve things that have only been enabled by massive amounts of computing capacity, that's that's sort of where they they that's why we sort of have them intertwined this week as technologies that have to do with the energy and the possible things of our future, the things that will affect our future in a massive way.
• 3:42
  So, absolutely. So the question I think we've decided to ask this week is what are these technologies? What is what is cryptocurrency?
• 3:50
  What is artificial intelligence and their broadest terms and their most specific terms?
• 3:57
  What is their true environmental cost? And will that energy demand act as a catalyst for innovation or negative consequences?
• 4:05
  Like you said, we basically fueling the dumpster fire that we've already seen coming about.
• 4:10
  Or are we saving ourselves from the dumpster fire by building the technology that's going to allow us to escape?
• 4:16
  Yeah. And our goal here, you know, obviously is not to to answer that question for you, but it's to sort of present the details and present the background to lead you to your own understanding or your own your own belief based on the facts of the situation rather than your understanding of what you think crypto is and what you think AI is and those sorts of things.
• 4:37
  Exactly. The one thing we won't do for you on the overcast overlap is we won't think for you.
• 4:42
  We'll talk to you. We'll research for you. We'll do a lot of work for you, but we're not going to think for you. We let you do that for yourself. Absolutely.
• 4:51
  Well, do you want to get started in those in defining those technologies then? Let's do it.
• 4:57
  All right, so we're going to kind of break down what each of these technologies are. I'm going to I'm being the tech, the tech lead here at the overlap.
• 5:07
  I'm going to going to kind of break it down, you know, obviously I'll start with cryptocurrency because it's probably the easiest to to understand and to grasp.
• 5:17
  Although it is more complicated than most people understand, I think there's a commonly held conception or not conception comprehension of cryptocurrency as just being a digital money right and and especially with with your average lay people out in the society.
• 5:34
  They know that it's a sort of an investment vehicle somehow that has some value and it could be speculative or not.
• 5:41
  So I want to kind of start talking about what what cryptocurrency actually is in its real life understanding.
• 5:51
  So ultimately, cryptocurrency, yes, is a system of money that is decentralized. So I do have to paint a picture of a central money system, but that's one that you're going to be very used to.
• 6:08
  So I'm not going to spend a whole lot of time on it, but ultimately what we have in in the United States now is what's called a fiat currency.
• 6:17
  There is a currency that is distributed by the government that's not backed by anything except the trust that the government uses that form of currency for the exchange of goods and services.
• 6:29
  Nothing fancy there, but inside of that system, we have a central bank. In our case, it's the Federal Reserve.
• 6:36
  The Federal Reserve keeps track on a ledger of all of the financial transactions that happen at our government level.
• 6:46
  So even though even though it is it's called centralized banking, it's kind of like an octopus or a squid.
• 6:54
  And I really like that analogy that will. So Will and I were kind of talking before the before the podcast started and he said, kind of like an octopus with tentacles everywhere.
• 7:02
  And I said exactly. Because typically when we think of a centralized system, everything happens at one place and then it's distributed outwards.
• 7:11
  But it really is a little bit different in the terms of our monetary system. So you as an individual, you have a bank or you transact money on things like PayPal and Venmo or inside of a banking account or something of that nature.
• 7:28
  So those transactions are actually kept on a local ledger by the transactional institution.
• 7:35
  It can pass through a lot of processing fees. It can pass through a lot of processing organizations like like Visa MasterCard and those sorts of organizations that their job is to help larger institutions exchange money with smaller companies or businesses or individuals.
• 7:53
  At that point, then the Federal Reserve actually comes in and and handles the bigger bank transactions. When I say bigger bank, I'm talking the big ones, Wells Fargo, JP Morgan Chase, you know, as well as a lot of the the financial institutions in New York City involving the stock market and the stock exchange.
• 8:12
  All of those transactions are settled on a giant ledger that is the Federal Reserve is kind of in charge of. And that's what they kind of do. The problem with that is the Federal Reserve, as we're seeing now, is is under sort of a political realm.
• 8:29
  It was always kind of a private bank and if you've if you've listened to any libertarians rattle on about it, it was never really designed like this. It wasn't until I think the 20s, 1929 or so that we had sort of a central bank that that played this giant role.
• 8:46
  I say 1929 I think I meant 1829. It's it's pretty old is the point that I'm trying to make. But at one point that we did not have it was Andrew Jackson. So whenever Andrew Jackson was alive, but Andrew Jackson basically allowed for a central bank to take root in the United States as that supreme ledger holder.
• 9:06
  It can be politicized. Like I said, there are appointees that are appointed by government officials. There is they have governors within their system. It's a pretty pretty secretive system and it hasn't been audited in a number of years.
• 9:21
  They kind of do an internal self audit, but it's very black box. You don't get to see inside of it. You don't get to understand their processes. You don't understand how or why they do the things that they do, but they also control the Fed rate.
• 9:35
  That's where you get low loan rates for things like banking and car loans and you know, large loans that are made to and from governments where, you know, like we have $37 trillion debt in the United States. Well, that has to be kept track of somewhere and they keep track of that debt on behalf of the United States as a government.
• 9:57
  So Fiat currency central banking system. We have the Federal Reserve Bank. That is the traditional money system that we're used to understanding in a decentralized system.
• 10:09
  There's something called a blockchain. What is a blockchain? It's a pretty simple collection of algorithms. It's an algorithm that comes together and produces a specific stream of data. And as transactions are transacted, it adds on to that stream of data.
• 10:28
  Once that stream has recorded that data and a transaction is on that ledger, the blockchain, it's permanent. It's fixed. It cannot be changed. It can be exchanged.
• 10:42
  ownership can be moved, but what is actually inside of that blockchain cannot be changed cannot be it's it's immutable. That is its strength. Now, because of that inside a cryptocurrency that blockchain and which I'm going to call a ledger from here on out is distributed amongst every single node or wallet.
• 11:10
  Not really a wallet, but every single node on that network, for example, Bitcoin or Ethereum and amongst those nodes can be anybody. If you if you want to purchase a node, a miner node as they call it, and put it on the blockchain, you will be able to download a full copy of that entire ledger.
• 11:32
  And that is how they keep track of transactions within this monetary cryptocurrency system.
• 11:41
  There are a couple of different sort of models for how they they determine the value of cryptocurrency. It's not agreed upon. It's debated wildly, and there's lots of theories. But first, I want to talk a little bit about the process of actually mining cryptocurrency, because they are two distinctly separate tasks that happen within the the cryptocurrency realm.
• 12:08
  So the first process is that sort of the mining process, and that is literally how cryptocurrencies are created.
• 12:17
  Basically, so basically, these miners are able to take in the blockchain and large pieces of data. Now this is done on purpose. In the case of most cryptocurrencies, because if your phone could do it right, for example of one person's phone could do it, then it would be a whole lot easier to fake this data.
• 12:37
  Right. So, ultimately, the computations that have to be done to validate and verify the data that's in the blockchain needs to be complicated. It needs to be difficult to crack or solve in this case, which is kind of where the power generation problem comes in, they're highly computational algorithms, they're highly computational problems that have to be solved by each of these mining computers.
• 13:05
  And these are these are not your everyday desktop computers there, you know, a long time ago, it was a little simpler, you actually could run a little bit of a Bitcoin mine on a raspberry Pi or something of that effect.
• 13:19
  And as the cryptocurrency has gotten more widely used, and more blocked out those problems become harder over time. And so we're talking about very specialized highly intensive graphic processing units that it takes to actually compute these transactions.
• 13:41
  Right. So it's essentially a peer to peer network, and that is specifically what makes them secure, it makes them traceable, and it's validated through that public ledger.
• 13:54
  So, if somebody says this person traded this coin for for this amount of fiat currency, and it goes into the blockchain, if it goes into the ledger. It's going to reach out through several people through several nodes within this network and validate that that transaction is correct.
• 14:14
  And once it's been validated that problem is solved.
• 14:18
  Then, that is considered a mining process that block, then goes into the blockchain. And in exchange for mining that block.
• 14:29
  The person who minded or the computer that mind it or the node that mind it receives cryptocurrency in exchange for mining that part of the blockchain.
• 14:40
  It's convoluted, I understand, but if you essentially think it, think of it as Bitcoin itself said, this is how big the blockchain could possibly ever be.
• 14:52
  Each computer node along here. So if I'm going to, I'm not going to give you the real numbers because it's something very large. It's but let's just say it was 100,000 right so if there were only ever 100,000 bitcoins that could ever possibly be mind, that's blocks on the blockchain on the ledger.
• 15:13
  There's only 100,000. So it starts out with the first one with easy problems to solve because the blockchain is smaller the ledger smaller, the ability to process and understand data from that ledger is smaller chunks of information smaller problems to solve.
• 15:30
  So it's worth less at the time, now, we're. Once we've got up into these these big miners and things of that nature.
• 15:40
  And by the time one block is mind, it gets put on the blockchain that means there are only 999 or 9999 9999 blocks left to be mind with this in this entire network so you have something that creates kind of a scarcity.
• 16:00
  Right, there's only a certain number that can ever exist.
• 16:05
  And this also introduces new cryptocurrency into circulation in a controlled manner, as opposed to just just being able to print money like we can with the Federal Reserve.
• 16:18
  It provides the computational work necessary to validate those transactions and then secure the network.
• 16:26
  So what we're doing is sort of user, the user interaction piece right so if you want to participate in a cryptocurrency ecosystem and for the case of this podcast we're going to use Bitcoin.
• 16:37
  It's easy it's prolific people know what it is and they've heard of it. So, it's involves a few key steps right so first you have to choose a platform.
• 16:47
  And this can be either like a broker that offers crypto assets, or it can be like a dedicated cryptocurrency exchange.
• 16:56
  I'm going to mention a couple of the larger ones. I believe there's Binance. I believe there is crypto.com Gemini is also one of them.
• 17:06
  That is not an endorsement of any of these platforms, I do, I personally use a couple of them. I don't like them.
• 17:13
  And I've chosen for the most part to keep the cryptocurrency that I do own on a cold wallet at my house.
• 17:21
  Because of some some issues I had with those platforms in the past so please do your own research on that if you do decide to use one of the cryptocurrency exchanges, but the, the, the idea behind the cryptocurrency exchange is that they will allow you to take your
• 17:40
  issued fiat currency like the US dollar, and then you can purchase your desired cryptocurrency in its amount, right. Well then once you buy that.
• 17:50
  It has to be stored in a digital wallet.
• 17:53
  These wallets don't actually hold crypto like when you get a Bitcoin wallet, it's not. There's not actual Bitcoin in this wallet. It's actually private keys, it's a it's a security key. That's very very complex very long, very complex cryptographic codes that basically grant you as the owner access to those funds that exists on the blockchain.
• 18:19
  And wallets come in two main forms hot wallets, which are software based and connected to the internet.
• 18:25
  You can actually download apps for your phone, you can, you can have a dedicated like software wallet that it has is Wi Fi capable or Bluetooth capable and you can control through an app on your phone as well.
• 18:39
  And there's also cold wallets and those are physical devices they look a little bit like USB drives.
• 18:45
  And they store those keys offline for security reasons.
• 18:50
  Now this is the storage method is is a, there's a fundamental trade off in the digital asset world between convenience and security. So you have to choose for you which is the best option, we're not going to endorse, which one is the best or the best I'll tell you what I do personally.
• 19:07
  I, after poor experiences on exchanges on hot wallets I chose a cold wallet for that very purpose. And that very reason. So, the next thing is is how how cryptocurrency gets its value, there are a couple of different theories behind how cryptocurrency
• 19:25
  specifically received its value. We're going to talk about a couple of those valuation methodologies. One, of course, is the cost of production model and this is probably the most widely known of your of the crypto bros, right, the crypto bros that are out there,
• 19:40
  hocking it and trying to get people to buy it. That's the most well known one. Ultimately, the belief is that they're that since most of the early crypto models are called proof of work cryptocurrencies, like Bitcoin we're talking about, basically says that the intrinsic value of Bitcoin is anchored directly to the costs required to produce it.
• 20:04
  What does that mean? That means the silicon to build those graphic processing units. That means the hardware and the labor and the transportation to get those mining platforms to validate those transactions on the cryptocurrency network, and also the power consumption and
• 20:22
  Obviously, you know we have a we have a slanted approach to how we're doing this podcast today and that's that's what we're kind of bent on is the power production. So that's the first step. They say if it costs $86,000 to mine one Bitcoin.
• 20:39
  That can be seen as kind of the baseline for its intrinsic value. So whenever they mine that specific coin or that specific block. If one Bitcoin is worth $86,000 this I think the spot price today is 111,773.
• 21:00
  So they they're up. I mean they would they would have made a profit if it costs $86,000 to do that because then they could get one Bitcoin. They could trade it in for fiat currency and receive $111,773 in cash minus fees, so that acts kind of as a floor for the price of the asset.
• 21:21
  That is again the cost of production model. There's also the network value and this is this is referred to as met cast law. I like to think of this in terms of like it like a toll road or a tollway.
• 21:37
  If you think of like a toll system in a highway system, you the toll company. Now I don't like tolls or toll roads or toll organizations. I think we should just build roads, but a toll company says look in exchange for us being able to collect.
• 21:55
  A fee for drivers to drive on this road. We'll build this road at our expense. And then once that's paid off and we've reached certain percentages of profit, we'll hand that back over to the city.
• 22:07
  Spoiler that actually never happens, but and we'll talk later about why, but that is that is the underlying cost of the toll road system. It's a collection of toll roads. All of them have an intrinsic value of base value that was put into producing those roads.
• 22:28
  Then you have the vehicles that actually drive on the road. So when they're paying $1 and 25 cents, or if you want to drive from Oklahoma City to Tulsa, it's going to cost you about $10, you know, on a toll road that $10 per car.
• 22:44
  On that along that toll system. That's going to be sort of the play against the intrinsic value of the network. So the value of the network itself is going to be both the intrinsic value of creating those highways, while also including the value of each car that passes along it.
• 23:08
  And an example of that is kind of like, is Ethereum. So you, you can you have a theorem as a blockchain, but you can actually build tokens and coins on top of that network called D apps right and build upon its blockchain.
• 23:25
  So the more D apps that are built, that's more traffic. So more traffic means the overall network is worth more. So, in Metcalfe's law, otherwise referred to as network value.
• 23:39
  That is the valuation. So if every single toll road, plus our average daily toll payers is included in this to whatever degree or time period that you're trying to to calculate the value for it, divided by the number of coins in this scenario that they stand for will be the the underlying network value.
• 24:06
  So not necessarily how much it costs to produce it, but not how much money how much fiat currency has been converted, and what is owned on it. At the same time, there are other models.
• 24:20
  There's also the discounted utility model. And that's sort of traditional financial valuation.
• 24:28
  They try to figure out the underlying assets, the future growth, the volatility and a delta.
• 24:35
  That is super boring.
• 24:38
  The two main ones that are used to produce value and the ones most touted are the two that I listed there, the proof of work model, as and the, or it's not the proof of work, but the
• 24:51
  cost of production model that the cost to produce it, it all combined is worth its value is what gives it value. And then the the entirety of the network itself, having intrinsic value.
• 25:06
  Like I said, they're, they're highly debated. I will say this. It is. It is all essentially speculation and people point that out to me a lot when they would like to argue about the benefit of cryptocurrency over another.
• 25:22
  I don't. I have both right like I had traditional savings account value things I have traditional cryptocurrency I have stocks bonds, you know things like that. But ultimately, I think it's a combination it's a it's part of an overall
• 25:38
  investment strategy. I wouldn't say, hey, I'm all in on Bitcoin or hey, I'm all in on Solana or Ethereum or whatever. So, well, can you think of anything that I left out there? I think that's that's an overall understanding of crypto how it's valued and why we value.
• 25:55
  Oh, I do want to mention before we go on that the value of the dollar that we use every day that we call Fiat currency is ultimately the same system. So when somebody points out that cryptocurrency is somehow inherently less valuable than the US dollar.
• 26:14
  I say, look, it really comes down to who accepts it, which is which is the real reason why we use the US dollar. It's not because we think it's the best. It's because when I take a US dollar down to a restaurant.
• 26:27
  I mean, it's not going to buy me anything, maybe a gumball nowadays, but it it's going to have intrinsic value to the person that is running that business right so its intrinsic value is only built on the trust the United States government and I will tell you worldwide right now that trust is slowly devaluing the US dollar.
• 26:48
  And by slowly, I mean very rapidly. If you check them, but the international markets out there. So yeah, I wanted to point that out the both of them are essentially just speculation.
• 26:59
  And I call it the greater fool right. It's all the greater fool method. It's the intrinsic belief that this is an investment because I can buy it at this price and later convince someone else a greater fool than myself to buy it at a better price.
• 27:19
  And that is true for both Fiat and crypto.
• 27:22
  Agreed. Now, talking sort of philosophically about this, like you sort of brought that up in the last point about the nature of value.
• 27:33
  It creates a lot of interesting problems and puzzles with all of these sorts of methods of exchanging value right.
• 27:41
  So one thing we didn't really cover there and it's not necessary for our discussion today in a sense, but it does kind of come back into play tangentially I think is that.
• 27:50
  So the alternative to a fiat currency most world currencies or fiat currencies these days, as I aware of them. And that was not always the case right prior to believe it was 1971.
• 28:01
  The US dollar was on the gold standard. That's right. The gold standard was a, and the thing about the gold standard right is that it basically had an anchor that there was this much gold, you know, in theory stored in the US government's treasury.
• 28:16
  Fort Knox is the famous repository of that and then based on the amount of gold. Allegedly right based on the amount of gold that was allegedly in Fort Knox or in another storage places that told you how many dollars you could have in circulation right.
• 28:33
  And that was the theory that it limited they gave it a finitude that it wouldn't otherwise have that you could just print money willy-nilly if you didn't have that limitation and of course that limitations gone now because we're off the gold standard and essentially.
• 28:47
  Maybe maybe it's willy-nilly maybe not depending on how you look at it, but the Federal Reserve has.
• 28:53
  The rear the expression, but a blank check to decide how many dollars are out there. Right. I mean they they have to follow certain procedures, but if they vote to increase the monetary supply.
• 29:03
  They literally print more dollars or in reality these days they print more or they move some numbers around on a digital ledger and say that there are now X number of dollars in circulation, whereas yesterday there were, you know.
• 29:17
  20 billion 20 trillion less than that.
• 29:22
  So that's one thing I think it's worth kind of noting is that you're right in the Fiat currency there's there's no limitation and the only valuation the only thing standing behind all those dollars is the promises of the US government.
• 29:34
  And you determine whether you value those or not in terms of exchanging US currency for other currencies other world currency.
• 29:41
  Yeah, and and there I mean there's actually been several there's been several backed currency standards in the United States eight 1792 with the coin a Jack they actually silver was created as the first standard for American currency and then they it sort of reached a peak.
• 29:59
  And then it was I think 1933 that Roosevelt was the one who started the phase out of the gold standard by confiscating some gold which was weird it was a very strange time in history if you if you're a history buff you should go check that out but 1963 was executive order 1111 zero by JFK on June 4.
• 30:27
  So that is what amended a previous executive order from 1951 that officially abolished the end of the gold standard.
• 30:36
  Right. And if you look there if you're interested in this stuff we won't nerd out on it too much.
• 30:41
  I'll try to resist anyway but there are some very interesting books. I know there's so many things like I feel so many tangents sometimes I just want to go down and I'm like oh yeah we have a time format.
• 30:52
  But, you know, there there are some very interesting books out there some supplementary reading maybe we'll add some links to our show notes on the books that are out there that you can go into to figure out the history of money because it really is kind of mindbending when you think about it you know money exists because of the friction between trading you know, barley for you know.
• 31:13
  Metals for your building your house. And if the if the blacksmith in town didn't need your barley, then what would you do you have to find somebody who needed barley to give you corn and corn to exchange for goats and you know just to take it as impossible right goats to exchange for your wife and exactly.
• 31:32
  All sorts of exchanges happening and so we decided that they'd be simple it'd be simpler to have some sort of way of keeping track of that. It's good for the exchange of the value the actual real items.
• 31:42
  Now it's basically in no way tied to any real item in existence, it's tied to some beliefs about the future productive capacities you know that the future contributions that your economy can make.
• 31:54
  And not much else, which is in a lot of ways, incredibly scary.
• 31:59
  Right, exactly. If you think Bitcoin or cryptocurrency is scary. Think about the fact that nothing but the future of your society determines the value of your currency on a given day.
• 32:10
  So, yeah, I think that's an interesting point and I kind of had this visual of what you're talking about this is like people in back rooms with ledgers and quill pins, marking down every transaction in the early days of the Federal Reserve, you know.
• 32:24
  And nowadays, I'm sure they have secured, you know, lines to from Wells Fargo's wherever they store their information directly to the Federal Reserve to upload all that sort of stuff.
• 32:35
  I'm sure it's much more complicated now but at the end of the day, it's really like, if you had access to the Federal Reserve's computing, your computer system or their whatever software they use to govern all that you could essentially have a license to print money.
• 32:50
  Yeah, which is if you if you manage to put the right bank information together, you absolutely could print your own money.
• 32:58
  Whereas my understanding of cryptocurrency, and please feel free to correct me on this because I'm sure I'll say something wrong.
• 33:04
  But you have sort of a mathematical limit on the number of coins, right. It's not it's not even like a it's a technical limit, I guess rather than mathematical limit.
• 33:13
  But it's a choice actually. It's a it is it is an instituted limit.
• 33:18
  Right. But you can do some supplemental reading on that with the bit bitcoins create around bitcoins creation was they did it to create it as an intrinsic value because if you could just create it forever, then it wouldn't have a fixed value.
• 33:34
  Right. And so essentially, I guess there's going to become a point where my understanding and I don't know if this was just a sample or example of it or if this is what they really do but something like calculating the prime factors of very large numbers was how they determine like the mathematical formulas used.
• 33:52
  Yes, and I would just as a side note on that I, I, at one point believed that that's what the nodes were doing.
• 34:01
  Like each node, all they were doing was finding infinitely more prime numbers. And I had this belief that that was the intrinsic value of crypto when I bought my first one.
• 34:11
  And so I hope I hope that you're out there and you're listening to this and you're like, Oh my god, I thought that exact same thing.
• 34:18
  And now I realized how dumb it was. I would love for you to leave a comment in the show at the show website or send us an email at the overlap at fof. Foundation, and tell us your story of what you believed cryptocurrency was.
• 34:35
  Excellent. Well that that's about the limits of my knowledge on cryptocurrency you've expanded my knowledge considerably today.
• 34:41
  Do you want to move on to AI or. Yeah yeah let's let's let's talk about AI, that one that was a little bit more complicated right. So, first off, I want to say this.
• 34:53
  AI stands for artificial intelligence. That is a gigantic umbrella for a lot of other things that people don't know the names for some are gaining relevance as the relevance of AI as an umbrella term increases things like GPT are becoming synonymous with AI.
• 35:17
  LL M's are becoming synonymous with AI. So, I think the first thing that we should do is, let's talk about AI as an umbrella right so as an umbrella it's a very big field of computer science that specifically deals with building machines and computer systems that perform tasks that would normally require people right like human intelligence, so they can include the ability to reason to learn from experiences.
• 35:46
  To learn from experience, understand and translate language, which is what GPT and LL M's are analyze complex data and act upon that information to make recommendations, decisions or predictions.
• 36:01
  There's been a large one, I believe it was on I don't know if was jeopardy or something it was the ENIAC computer that was trained specifically on how to play chess.
• 36:14
  It could be grandmaster so that is actually AI I mean it's so AI in and of itself is just predictive models, and it's that it encompasses this umbrella of training a model, and then creating a model that takes input to create prediction.
• 36:38
  So, the first scope, the first thing, term that you can bring to your to your bridge club game, depending on how old you are, I don't even know what a bridge club is.
• 36:47
  I know it's a game with cards but bridge. No, so you can, you can take this to your grandpappy and talk to him about it if you want to artificial narrow intelligence, a ni.
• 36:58
  It's, it's, it's important to first understand that all AI that are in widespread use today are actually artificial narrow intelligence. What does that mean.
• 37:11
  It means that there are systems that are designed and trained to perform a specific set of narrow tasks. So, a Google search algorithm right is an example of that a speech recognition system that takes our audio and I use it every every time we put out a podcast that transcribes
• 37:31
  our audio into text right, or an AI that looks for product defects on a factory floor are all really powerful examples, a chess computer is perfect examples of artificial narrow intelligence, they perform their specific function with faster than human speed
• 37:52
  faster than human accuracy, but their intelligence intelligence if you want to use that word is really confined to that thing. That specific thing that was trained on. So, if you were to take that EDM computer and take an AI that's trained to play chess.
• 38:10
  It can't give you medical advice, or teach you how to drive a car. Right. And that's, that's a very important distinction.
• 38:19
  Because they can only stay in its lane.
• 38:22
  Right, exactly. So, so let's, let's talk a little bit about about the one that's been kind of the buzzword around and that is chagy PT right so open AI really burst on the scene with chagy PT as the first, and we're seeing so many come around now, there's Claude,
• 38:39
  there's,
• 38:42
  with Gemini through Google, you've got Siri now that's serious Siri AI enabled. You even have independent models through llama and Facebook. Now you can use Facebook chat to talk to the meta AI.
• 38:55
  Those are actually language learning models, LL M's. So, they're not GPT is chagy PT is a is a brand name for open a eyes, LL M. Gemini brand name, Google's LL M language learning models.
• 39:15
  Sorry, large language models, I always get that messed up actually let not language learning models, large language models. So, they are a type of generative AI that specializes in understanding translating processing and and creating human like text.
• 39:39
  So models like open a eyes GPT or Gemini are trained on gigantic internet scale data sets of text and code.
• 39:50
  Almost every single AI is trained on the entirety of Wikipedia at a, at a bare minimum for for large language models. In addition to GitHub and get lab and code repositories that are out there like like stack trace and those sorts of things.
• 40:11
  The way that they work is they analyze all of these trillions of words, and then learn the statistical patterns the relationships, the grammar, the, the spaces between how many words are between these, these statistical patterns of human language, and it's, it's not just limited to one it's limited to all.
• 40:32
  So their core function, really, is to is highly advanced prediction.
• 40:41
  So if you give a little LL M a sequence of words. It calculates the most probable next word, and then the word after that and then so on. And that allows the people who are running these models to compose coherent contextually relevant sentences paragraphs and sometimes whole documents
• 41:04
  spreadsheets and you know all the things that AI or LL M's can produce I do it myself, I say I all the time. And we say on the show and we absolutely mean it in place of large language models, but that's what it is.
• 41:19
  Well, generative AI I've used this this term a lot so generating something so creating something generative AI versus large language models can be kind of confusing. I think that it's distinctly the distinction between them is that generative AI is the next step of the
• 41:42
  formula of AI so we have AI, a NI generative AI. So under that you have large language models you've got things like diffusion models that generate create pictures and photos and videos and all these other things underneath them.
• 41:59
  So, even though generative AI is kind of the broad category for any AI system that can create original content and LL M's generate text right or other forms of generative AI can produce songs or mid journey and Dolly have massive data sets of images and that generate pictures
• 42:22
  from text, other models generate music and video and synthetic data for scientific simulations.
• 42:28
  AIs are now currently being used to generate data to train AIs.
• 42:35
  Or to train train other models.
• 42:39
  All of those are forms of generative AI, but not all generative AIs are large language models like all squares or rectangles and not all rectangles or squares.
• 42:50
  Exactly. So the next step in that is is understanding the when we talk generally about AI, and our fears and our hopes and our aspirations and our desires. What we're actually talking about is a GI artificial general intelligence.
• 43:11
  And so, a GI to date is a hypothetical form of AI. It doesn't actually exist yet. Now, all of these things generative AI LL M's GPT. The visual aspects of generally AI are all baby steps that will eventually lead and have to take place that lead to a GI.
• 43:34
  So what is a GI theoretically, it's an intelligent system designed to solve any problem within the breadth and proficiency of a human being. So, that means, once we have a GI, it will not only tell me how to how to make a new recipe.
• 43:54
  It can show me how to drive, it can detect anomalies in manufacturing processes, it can drive a boat.
• 44:02
  And then it can build a starship you know what I'm saying like and then even go further and say you know even though humans right now can't really build a star, a starship to move in between stars.
• 44:13
  The hope that some of us have will and I have talked about it before is that eventually once we have a GI, they will be able to feed off of each other and create an ASI. So that supersedes general intelligence of humanity.
• 44:32
  Ultimately, what's important to know about a GI is that they're, they don't need to be programmed to a specific task they're programmed to do anything within the human realm.
• 44:43
  Obviously if it's a computer model it's going to be limited to, you know, the meat space.
• 44:49
  But LL M's are not a GI.
• 44:53
  You're probably gonna be wrong. They have remarkable fluency in multiple languages, including computer languages.
• 45:00
  It's just really important to recognize that current LL M's are not actually a GI at their core.
• 45:07
  Their spicy predict a text.
• 45:10
  And that's it. They're predictive text machines. They don't understand concepts in the human sense of the word.
• 45:18
  They actually process and manipulate statistical relationships between words and phrases that they were able to ingress from their training data.
• 45:30
  So yes, they can generate a poem about a cat.
• 45:34
  But it doesn't actually have genuine knowledge of what a cat is or even what a poem is. It comes from having analyzed countless examples of text about cats and poems and learning the probabilistic patterns of how words are combined together around those relationships.
• 45:57
  In contrast, an AGI would understand why speed limit signs exist.
• 46:05
  Currently, we have, we can ask about speed limits and it will give us the what a speed limit is, right?
• 46:13
  And tell us what that actual number is and then generate a photo of a speed limit sign.
• 46:19
  But it doesn't actually understand why there are speed limits.
• 46:25
  You might be able to ask it, why are there speed limits and it will give you an answer that is going to be a combination of other people's thoughts and other people's ideas.
• 46:35
  Yeah. So, an LL M would know from its training data that images of the stop sign or the speed limit sign are correlated, because it matches the visual representation of a sign and what signs are believed to be understood.
• 46:51
  And just can describe cars slowing down and the LL M's performance is really just an imitation of understanding.
• 47:00
  But it lacks the actual comprehension.
• 47:05
  So, it's been debated that there are now sparks of an AGI, right? Like we see these little blips that people say, oh, look, it's becoming conscious.
• 47:21
  No, it's just, it's just unfortunately untrue. Now I say unfortunately because it'd be great.
• 47:27
  And we statistically have created tests like the Turing tests, right?
• 47:33
  Alan Turing created a test that basically said that a truly sufficient artificial intelligence could hold a conversation that was indistinguishable from that of with a human, right?
• 47:50
  And look, GPT, Cod, Gemini, they're becoming really good at generating human like text that are actually getting closer than ever actually passing the test.
• 48:04
  That intrinsically means the test is bad. We need to create, and we have a responsibility to create a new test, right?
• 48:12
  But the research behind all of this is is that the imitation of intelligence is not intelligence.
• 48:24
  So, an LL M is just a really fancy advanced guessing engine, not a thinking mind.
• 48:33
  And there's something called being called the Turing test trap. And the trap is that because we've poorly defined the test, we're creating an ANI that passes the test in the same way that we create students that passed aptitude tests.
• 48:51
  Yeah, so the Turing test gets pretty complicated here if memory serves correctly, because Turing was actually sort of he was operating on this idea that there was what he was named after him eventually as the universal Turing machine, which was basically this sort of loosely defined device that could come compute anything basically.
• 49:12
  And if you had been one of those sufficiently complex he believed that it could hold a conversation with you via text and he had this whole experiment that he designed.
• 49:22
  But the problem now is that we know how, to some extent, although I think you'll get into this maybe a little bit in a minute, but there's a, to some extent, the LL M's, the various ANIs that we have access to now are a bit of a black box to us, we understand them to some degree
• 49:41
  we understand that in theory how they work we don't exactly know entirely what's going on inside of one. Is that fair to say?
• 49:47
  Yeah, it's actually magic, even from the people who train them and who like who study this as an actual like thing.
• 49:56
  Right, so in a weird way, you have to be careful with the sort of popular understanding of the Turing test and Turing's actual test and then some other complications, which we unfortunately can't go into in extreme detail.
• 50:09
  But the idea is that the black box is kind of what Turing was saying was potentially the intelligence, even though we know that the ANIs that we have now like LL M's.
• 50:20
  And I'll try to use these terms a little bit more carefully than I normally do because of what we're talking about.
• 50:24
  But we're going to go back to the way that we spoke before I promise.
• 50:28
  Right, so in LL M, even though it doesn't think like a human brain thinks across multiple domains and using multiple modes of input and output, they do in some sense contain a limited set of intelligence or limited subset of our human intelligence in the form of our language.
• 50:45
  By knowing, so to speak, or by having weights assigned to the relationships between tokens or words, they do sort of contain the knowledge that's contained in human language.
• 50:57
  It doesn't mean that they know anything in the sense that we think of know, but there is an intelligence inside of them.
• 51:03
  And I think what Turing was trying to point out was that at some point when that becomes sort of that when that black box is unknown to us, his his whole point I think was just that we can't say that it's not thinking.
• 51:17
  It is more like you can't rule out the possibility that it's thinking given its inputs and outputs.
• 51:22
  If it gives you the right outputs to the right inputs to certain inputs, then it's said to be thinking in all the ways that matter for thinking, at least in terms of language.
• 51:32
  So that's that's the difference in I think what Turing was was ruling out versus what he wasn't saying that that suddenly conscious right that it's suddenly like, if you had a complicated enough.
• 51:42
  LLM it could become conscious. That's not how LLMs work. We know that now Turing didn't have the benefit of that knowledge, but I think he was limiting his statement to saying if you don't know exactly what's going on inside that black box.
• 51:54
  And you put in an input and it gives you the correct output to that over the length of a conversation, you can't rule out that it's not thinking in some way.
• 52:03
  Does that make sense? Yes, yes it does.
• 52:07
  I mean, I want to I'll expand on that just a little bit, right? So the Turing test was actually I say invented, but he was originally called the imitation game, which is why the movie imitation game.
• 52:21
  But it was it was really sort of yeah, it was thought up by Turing in 1949.
• 52:28
  We're talking like way before computers, you know they were still working on the German code machine. But he actually introduced it in 1950 in a paper called computing machinery and intelligence when he was working at the University of Manchester, not Iowa University of Manchester in England.
• 52:47
  So it opens with the words, I propose to consider the question, can machines think, because thinking is difficult to define though he chose to replace the question by another which is closely related to it and is expressed in relatively unambiguous words.
• 53:07
  And so Turing actually described the new form of the problem in terms of a three person party game called the imitation game.
• 53:15
  So the interrogator asked questions of a man and a woman in another room, in order to determine the correct sex of the two players. So the question is, are there imaginable digital computers, which would do well in the imitation game.
• 53:30
  So the question, and because he believed it, was one that could actually be answered like it could, we can we can think right. And so the remainder of the paper he actually just argued against the major objections to the proposition that machines can think.
• 53:47
  So, actually, we have, we have developed the Turing test. And really so it's really just to answer the question, can machines think. And so we are can we say we, we as a as a humanity and people who are working to on AI and toward language learning models and toward AGI.
• 54:05
  We collectively the Royal we are are continually asking that question. And all of our tests are designed to ask that question. Well, to, to replace the question, can machines think.
• 54:17
  Yes, in a sense, it's the answer the question, can they do something that we would think, by all accounts requires thinking, as opposed to because thinking is not so we'll define right, we have a term that we all use every day that we can actually define so he's like let's define it via actually just thinking that as you were saying it I was like, can we even do we even know what thinking is. Yeah, this is the stuff of many philosophy courses.
• 54:43
  Exactly. Well, and that's that's the thing right like that's
• 54:49
  that idea has created this whole philosophy around AI. And it's why it's why in my opinion we have to regulate AI a lot better than we are currently doing.
• 54:59
  I think just last week. There are some parents who unfortunately their their child unalived himself. It was, it was a really bad situation I read the transcripts, he was talking with Chagy PT at the time, and a lack of understanding a lack of comprehension from the the language learning model, our large
• 55:21
  language learning model, created a scenario where it actively helped him cover up his attempts, and to encourage him to hide a noose that was he was hanging in his room in the hopes that somebody would see it and try to help him and intervene.
• 55:37
  So, that's why these things are important, right.
• 55:42
  They take up a lot of electricity. Let's talk, let's talk about that that big flaw in the, and the energy footprint so before we get into the, the nitty gritty details of it I do want to talk a little bit about the energy footprint of both cryptocurrency
• 56:03
  and AI.
• 56:06
  So, cryptocurrency has been around a lot longer, really since like the 2006 to 2008 kind of area so we have a lot more data. We have a lot more background information, Bitcoin just Bitcoin by itself.
• 56:23
  It consumes between 127 and 172 terawatt hours of electricity annually.
• 56:31
  Now, if we want to put this in perspective, because I don't know about you. I don't think in terawatts, and none of that means anything to me like how many light bulbs get it power on no Norway, as a country consumes 124 terawatts of electricity every year,
• 56:49
  and Poland is right behind it at like 118.
• 56:55
  Yeah, so crypto Bitcoin specifically uses more electricity than Poland, and somewhere in the realms of Norway, as a country every year around the world.
• 57:08
  Now it's important to consider, it's a distributed network, there are nodes all around the planet. There's no way to specifically measure it, but that is, that is the, the current consumption based understanding.
• 57:24
  You can also consider that cryptocurrency miners actually seek out the cheapest electricity options possible.
• 57:33
  What does that mean.
• 57:35
  So, that essentially means that they're going to employ multiple things sometimes it's it's solar panels with a battery setup. That's eco friendly, maybe they're using wind generators, maybe they're relocating to a particular part of Nevada right
• 57:53
  now. So, if you're, if your entire setup is going to be solar with battery. It makes sense to go out to Nevada. And some place that's got sunshine year round and set up your solar panels that way.
• 58:05
  But if that means moving your entire Bitcoin mining operation to Detroit, Michigan, and take advantage of lower than average electricity rates in a poor area, or in Texas where they actively encourage businesses to steal and rape their resources, then they're going to go there.
• 58:34
  And and what that essentially means and you I've watched even some, there's a couple more perfect union videos on YouTube that I was watching about this, and they're giant Bitcoin mining operations that are put up in small towns and it's producing a couple of things.
• 58:50
  First of all, it requires greater electricity generation.
• 59:05
  So it's not. It's not a great thing. Right. So, and that's just here in the United States, but the economic reality drives crypto miners to constantly seek out the cheapest sources of electricity available anywhere in the world.
• 59:26
  And that is that is the overall cryptocurrency. Now, AI is a little newer, right? I mean, the extensive use that we have. I mean, it's, it's pretty rough in terms of both water and electricity, but we're seeing it proliferated everywhere.
• 59:41
  Everybody knows about chat GPT or, or some version of their preferred chat model. And I even had a friend and a listener of the podcast say that that he was learning perplexity now that flux had just been been released inside of perplexity.
• 59:59
  And I was just like but but why though and he's like, well, you got to get it get with AI if you want a job for the future. He's not wrong. I mean, it's there but that has a real world back end cost that companies like open AI and companies like like grok and companies like Jim and I have actively hidden from us, right?
• 1:00:26
  AI, the cost of AI in the training phase is fixed. So, there is there is a contingent that kind of says okay AI from an energy footprint perspective is inside of this inference phase it's the training phase that has a fixed cost, it will not be replicated.
• 1:00:48
  So even if that, and the art their argument is, even if the costs of the electricity at the initial training phase is high.
• 1:00:57
  That's not necessarily going to translate out into the cost of running that model. Now,
• 1:01:05
  we'll see the the other is obviously like I said, in is in running that model and the inference tasks that will well surpass the initial training costs and now you're seeing companies, creating new models on a regular basis and updating their old models, and there's a lot of reasons for that right new words that are created new sentences
• 1:01:29
  that are created throughout the world and we want. I say we again in the Royal sense, we want that model to be able to infer the newest collection of words and the newest relationship of other words we want that spicy predict the text to understand the date and understand
• 1:01:49
  the information relevant to as fast as the data is being produced in a very content rich and driven world.
• 1:02:00
  I think, I think I'm going to note there that you know it's if we're going to continue to use language learning models, the energy consumption alone is going to far outweigh the training costs.
• 1:02:12
  Right, because it's not fixed at some point there's going to be an amount of use that surpasses the training costs.
• 1:02:17
  Exactly. And I mean I would say that that you know like for oh, has probably far surpassed of chat GPT has far surpassed enough that people were upset that five Oh was apparently boring and not not as spicy as four oh girlfriends and boyfriends were, and so they had to even bring it back and dig it back up from that past.
• 1:02:39
  Right. Which are our use is going to determine the value we extract from that, you know whether we can use that the energy that's plugged into GBT and LMS will actually give us something back more than we plug into it which is our big question for this podcast right.
• 1:02:56
  I do want to mention to sorry before before you, you move on here. I do want to mention that there are some additional costs to both cryptocurrency and AI. And those are actually like water and carbon emissions right so there's lots of carbon emissions that are created by Bitcoin, because of cooling.
• 1:03:20
  And ultimately, all of that generation in both cryptocurrency and an AI generates a whole heck of a lot of heat. And so, those have to be consistently cooled which uses water which uses other forever chemicals in order to drive that.
• 1:03:36
  Right and that's waste heat currently right that's generated from the actual computational activity of the chips and the various processors and things like that.
• 1:03:46
  Yeah, I waste heat, it immediately makes me defensive I don't know why but I would say yes in some cases but in some cases that some of that he has been actually recycled and used to capture other processes or assist with other things.
• 1:04:03
  Like that's actually, I think it would be. Yeah. Okay, yeah, yeah, I think it would be really cool to see those heat those that actual heat being channeled into boilers that generate more electricity.
• 1:04:14
  That's exactly what I was gonna suggest is it would basically need to bury these processors and inside of turbines and turn them into steam, you know, turn them into boilers it'll turn water into steam and utilize that to turn.
• 1:04:25
  Unfortunately, it's still going to be a net loss, it will just slow the net loss down. The laws of physics right I mean the universe is eventually going to suffer heat death spoilers.
• 1:04:34
  Right, most likely, but.
• 1:04:37
  And not to mention in the mining sense, even though, you know, even though you can set up where the energy is the cheapest if the energy becomes cheaper somewhere else that will eventually be a tipping point.
• 1:04:48
  So then you've got to deal with the co2 emissions of of shipping it via railroad way and you know shipping it by shipping those miners by a truck.
• 1:04:56
  And in some cases, these are entire containers 40 foot by eight foot by eight foot containers, filled with these Bitcoin rigs.
• 1:05:05
  And they, they can put them on a truck and move them to where the electricity is cheapest than any given given month, as long as the cost doesn't out outweigh the benefit.
• 1:05:15
  Right.
• 1:05:18
  So, why the hope that these two technologies, although they're very energy hungry we understand now why they're so energy hungry.
• 1:05:27
  How do we come up with this idea that perhaps they could rescue us from the energy scarce or energy consuming future that we fear, and actually give us a positive way out.
• 1:05:36
  That brings us to the end of the first episode of this two part episode, make sure to check back next week for part two. In the meantime, get your overlap podcast fixed at our website at Fof foundation, or check us out on blue sky and mastodon as overlap podcast.
• 1:05:51
  See you next week.
• 1:05:52
  Okay.
• 1:06:19
  (upbeat music)
• 1:06:21
  [MUSIC]